Kyle Wiggers Shares Allen AI Benchmark Audit

Kyle Wiggers posted a link to an Allen AI status update. The post describes new evidence that LLM benchmarks are not measuring what many assume. A generated headline in the packet states that an Allen AI audit shows LLM benchmarks mix safety and reasoning signals. The accompanying source summary notes that Allen AI researchers applied the BenchMIRT tool to popular LLM evaluation benchmarks and uncovered details about those signals. The packet attributes the original comment to Wiggers in his role as creator and communications lead. No further outcomes or independent confirmations appear in the supplied lines.